TRANSFORMER CONDITION ASSESSMENT USING ARTIFICIAL

INTELLIGENCE

A THESIS IN ELECTRICAL ENGINEERING Master of Science in Electrical Engineering

Presented to the faculty of the American University of Sharjah College of Engineering in partial fulfillment of

the requirements for the degree

MASTER OF SCIENCE

by REFAT ATEF GHUNEM

B.S. 2008

Sharjah, UAE May 2010

 

 

 

 

 

 

 

 

 

 

 

 

© 2010

REFAT ATEF GHUNEM

ALL RIGHTS RESERVED  

 

We approve the thesis of Refat Atef Ghunem Date of signature _________________________ Dr. Ayman El Hag Assistant Professor, Electrical Engineering, AUS Thesis Advisor

__________________

_____________________ Dr. Khalid Assaleh Associate Professor, Electrical Engineering, AUS Thesis Advisor

__________________

__________________________ Dr. Rached Dhaouadi Associate Professor, Electrical Engineering, AUS Graduate Commnittee

__________________

__________________________ Dr. Michel Bernard Pasquier Associate Professor, Computer Science and Engineering, AUS Graduate Commnittee

__________________

__________________________ Dr. Mohamed El-Tarhuni Associate Professor, & Department Head Coordinator, Electrical Engineering Graduate Program

__________________

__________________________ Dr. Hany El-Kadi Associate Dean, College of Engineering

__________________

__________________________ Dr. Hany El-Kadi Associate Dean, College of Engineering _________________________ Mr. Kevin Lewis Mitchell Director, Graduate & Undergraduate Programs & Research

__________________ __________________

TRANSFORMER CONDITION ASSESSMENT USING ARTIFICIAL

INTELLIGENCE

Refat Atef Ghunem, Candidate for the Master of Science Degree

American University of Sharjah, 2010

ABSTRACT

As a result of the deregulations in the power system networks, utilities have been

competing to optimize their operational costs and enhance the reliability of their

electrical infrastructure. So, the importance of implementing effective asset

management plans that improve the life cycle management of the electrical equipment

is highlighted. From an asset management point of view, commonly practiced

maintenance strategies are considered to have large portions of redundant costs.

Therefore, applying cost-effective, reliable and conditionally-based maintenance

policy is a priority. Having certain and comprehensive condition assessment of the

electrical equipment supports the selection of the appropriate maintenance plan.

Power transformers are the most costly electrical infrastructures; hence, condition

assessment of power transformers is a necessary task. It is a well accepted fact that

the remnant useful life of the transformer paper insulation determines its useful

operational life. Thus, reliable and economic transformer’s insulation condition

monitoring and diagnostic techniques are necessary to conduct a comprehensive and

efficient transformer condition assessment.

In this dissertation, artificial neural network is utilized as a modeling tool to

predict transformer oil parameters. Accordingly, the diagnosis efficiency of several

transformer condition monitoring techniques are enhanced and both corrective

iii

maintenance and end-of-life assessment costs are reduced. The research is focused in

predicting parameters able to diagnose both transformer insulating oil and its paper

insulation condition. Transformer insulation resistance parameter is used as an input

for an artificial neural network prediction-based model to estimate transformer oil

breakdown voltage, water content and dissolved gases. Moreover, furan content in

transformer oil is predicted using artificial neural network with step-wise regression

as a feature extraction tool.

An effective prediction model of oil breakdown voltage, water content, dissolved

gases and carbon dioxide to carbon monoxide ratio with prediction accuracies of 97%,

85%, 88% and 91% respectively is achieved. The excellent prediction accuracies

achieved reduces inspections’ time of unplanned outages. Furthermore, Oil quality

parameters and dissolved gases are verified to be statistically significant inputs for the

correlation with transformer oil furan content. The correlation is confirmed with a

prediction accuracy of 90%. By achieving such accuracy, assessing the transformer’s

solid insulation and ultimately verifying its useful remaining life is approached.

 

 

  

 

 

 

 

 

 

 

iv

TABLE OF CONTENTS

ABSTRACT ............................................................................................................... III

TABLE OF CONTENTS ........................................................................................... V

LIST OF FIGURES ................................................................................................. VII

LIST OF TABLES ................................................................................................. VIII

LIST OF ACRONYMS ............................................................................................ IX

ACKNOWLEDGEMENT ........................................................................................ XI

CHAPTER 1 ............................................................................................................... 1

INTRODUCTION........................................................................................................ 1

1.1 TRANSFORMER ASSET MANAGEMENT .................................................................. 1 1.2 TRANSFORMER MAINTENANCE PLANS ................................................................... 2

1.2.1 Corrective Maintenance ................................................................................. 2 1.2.2 Preventive Maintenance ................................................................................ 3 1.2.3 Reliability-Centred Maintenance ................................................................... 4

1.3 TRANSFORMER CONDITION ASSESSMENT .............................................................. 4 1.4 TRANSFORMER INSULATION SYSTEM .................................................................... 6

1.4.1 Transformer oil .............................................................................................. 6 1.4.2 Transformer Solid Insulation ......................................................................... 7

1.5 TRANSFORMER INSULATION ASSESSMENT TECHNIQUES ....................................... 8 1.5.1 Insulation resistance test (Megger test) ......................................................... 8 1.5.2 Oil quality tests .............................................................................................. 9 1.5.3 Dissolved gas-in-oil analysis (DGA) ........................................................... 10 1.5.4 Tensile Strength (TS)................................................................................... 12 1.5.5 Degree of Polymerization (DP) ................................................................... 12 1.5.6 Oil furan content .......................................................................................... 13 1.5.7 Artificial Intelligence Applications in Transformer Condition Monitoring and Diagnostic………………………………………………………………… .. 14 1.5.8 Summary………. ......................................................................................... 17

CHAPTER 2 ............................................................................................................. 18

OBJECTIVES AND CONTRIBUTIONS OF THE RESEARCH ........................ 18

2.1 OBJECTIVES. ........................................................................................................ 18 2.2 CONTRIBUTION AND SIGNIFICANCE OF THE RESEARCH ....................................... 19 2.3 RELATED PUBLICATIONS ..................................................................................... 20

CHAPTER 3 ............................................................................................................. 21

MATERIALS AND METHODS .............................................................................. 21

v

3.1 TRANSFORMER INSULATION TESTS ..................................................................... 21 3.1.1 Megger Test ................................................................................................. 21 3.1.2 Oil Quality Tests .......................................................................................... 23 3.1.3 Dissolved-Gas-in-Oil Analysis (DGA) ....................................................... 23 3.1.4 Oil Furan Content ........................................................................................ 24

3.2 ARTIFICIAL NEURAL NETWORKS (ANN) ............................................................. 24 3.2.1 ANN Structure Proposed ............................................................................. 24 3.2.2 Feature extraction using stepwise regression .............................................. 28 3.2.3 ANN Training and Testing .......................................................................... 31

CHAPTER 4 ............................................................................................................. 34

RESULTS AND DISCUSSION ................................................................................ 34

4.1 OIL QUALITY PARAMETERS AND DGA PREDICTION MODEL ............................... 34 4.1.1 Oil BDV Prediction ..................................................................................... 34 4.1.2 Oil water content prediction ........................................................................ 37 4.1.3 DGA prediction ........................................................................................... 38

4.2 FURAN CONTENT PREDICTION MODEL ................................................................ 42 4.2.1 Model Input Features ................................................................................... 43 4.2.2 Oil furan content prediction ......................................................................... 46

CHAPTER 5 ............................................................................................................. 50

CONCLUSIONS AND RECOMMENDATIONS ................................................... 50

5.1 PREDICTION OF OIL QUALITY PARAMETERS AND DISSOLVED GASES .................... 50 5.2 PREDICTION OF FURAN CONTENT ........................................................................ 51 5.3 RECOMMENDATIONS FOR FUTURE WORK ............................................................. 52

REFERENCES ........................................................................................................... 53

VITA............................................................................................................................ 57

 

vi

LIST OF FIGURES

FIGURE 1: SCHEMATIC DIAGRAM OF MEGGER TEST (HV/GROUND) .............................. 22 FIGURE 2: SIDE VIEW OF PRACTICAL MEGGER TEST CONDUCTED IN THE SITE ............... 22 FIGURE 3: ARTIFICIAL NEURAL NETWORK STRUCTURE PROPOSED FOR PREDICTING BDV, WATER CONTENT, TDCG AND CO2/CO ........................................................................ 26 FIGURE 4: ARTIFICIAL NEURAL NETWORK STRUCTURE USED FOR PREDICTION OF FURAN

CONTENT ....................................................................................................................... 26 FIGURE 5: STEPWISE REGRESSION FLOW CHART. ........................................................... 30 FIGURE 6: CORRELATION BETWEEN INSULATION RESISTANCE AND OIL BDV. .............. 35 FIGURE 7: CORRELATION BETWEEN INSULATION RESISTANCE AND OIL WATER CONTENT....................................................................................................................................... 37 FIGURE 8: CORRELATION BETWEEN INSULATION RESISTANCE AND OIL TDCG. ............ 39 FIGURE 9: CORRELATION BETWEEN INSULATION RESISTANCE AND OIL CO2/CO

RATIO…………………………………………………………. ................................... 41 FIGURE 10: FURAN CONTENT PREDICTION IN TRANSFORMER OIL WITH CO, TDCG, ACIDITY AND BDV AS INPUTS ....................................................................................... 46 FIGURE 11: FURAN CONTENT PREDICTION IN TRANSFORMER OIL WITH CO2, ACIDITY

AND BDV AS INPUTS ..................................................................................................... 47 FIGURE 12: FURAN CONTENT PREDICTION IN TRANSFORMER OIL WITH CO, CO2, ACIDITY

AND BDV AS INPUTS ..................................................................................................... 47 FIGURE 13: FURAN CONTENT PREDICTION IN TRANSFORMER OIL WITH CO, ACIDITY AND

BDV AS INPUTS ............................................................................................................ 48 FIGURE 14: FURAN CONTENT PREDICTION IN TRANSFORMER OIL WITH TCG, ACIDITY

AND BDV AS INPUTS ..................................................................................................... 49

 

 

 

 

vii

LIST OF TABLES

TABLE 1: MAINTENANCE TECHNIQUES EFFECT ON COST, PERFORMANCE AND RISK ........ 5 TABLE 2: IEC 60422 RECOMMENDED ACTION LIMITS FOR OIL BDV AND WATER

CONTENT TESTING [12] ................................................................................................... 9 TABLE 3: ACTIONS BASED ON TDCG ACCORDING TO C57.104 [13] .............................. 11 TABLE 4: WESTINGHOUSE GUIDELINES ON TCG [13] ................................................... 12 TABLE 5: SAMPLE OF THE DATA USED IN THE FIRST STUDY OF PREDICTING BDV, WATER

CONTENT, TDCG AND CO2/CO RATIO USING INSULATION RESISTANCE MEASUREMENTS

AS INPUT FEATURES ....................................................................................................... 23 TABLE 6: SAMPLES OF THE OIL DATA USED THE SECOND STUDY OF PREDICTING FURAN

CONTENT IN TRANSFORMER OIL .................................................................................... 24 TABLE 7: MEAN SQUARE ERROR AND THEIR RESPECTIVE NUMBER OF EPOCHS FOR

SAMPLES OF THE ANNS PROPOSED TO PREDICT BDV, WATER CONTENT, TDCG, CO2/CO AND FURAN CONTENT ..................................................................................... 33 TABLE 8: PREDICTION ACCURACIES OF BDV FOR DIFFERENT SIZES OF TRAINING SETS . 36 TABLE 9: PREDICTION ACCURACIES OF OIL WATER CONTENT FOR DIFFERENT SIZES OF

TRAINING SETS .............................................................................................................. 38 TABLE 10: PREDICTION ACCURACIES OF OIL TDCG FOR DIFFERENT SIZES OF TRAINING

SETS .............................................................................................................................. 40 TABLE 11: PREDICTION ACCURACIES OF OIL CO2/CO RATIO FOR DIFFERENT SIZES OF

TRAINING SETS .............................................................................................................. 42 TABLE 12: SELECTED TERMS OF DIFFERENT TESTED STEPWISE REGRESSION MODELS ... 43 TABLE 13: RECOMMENDED MODELS BY STEPWISE REGRESSION.................................... 44

 

 

 

 

 

viii

LIST OF ACRONYMS

AM - Asset Management

RCM - Reliability-Centered Maintenance

TBM - Time-Based Maintenance

CBM - Condition-Based Maintenance

CM - Condition Monitoring

H2 - Hydrogen

CH4 - Methane

C2H4 - Ethane

CH6 - Ethylene

C2H2 - Acetylene

CO - Carbon Monoxide

CO2 - Carbon Dioxide

IR - Insulation Resistance

DGA - Dissolved Gas-in-oil Analysis

BDV - Breakdown Voltage

TDCG - Total Dissolved Combustible Gases

TCG - Total Combustible Gases

ppm - Part per million

TS - Tensile Strength

DP - Degree Of Polymerization

ANN - Artificial Neural Network

MLP - Multi-Layer Perceptron

RAD - Relative Aging Degree

HV - High Voltage

ix

LV - Low Voltage

TV - Tertiary Voltage

SSR - Sum of Squares Regression

SST - Sum of Squares Total

 

 

 

 

 

 

 

 

 

 

x

xi

ACKNOWLEDGEMENT

By completing this research work, first and foremost, my gratitude goes to our

almighty God for giving me the faith, strength and hope to complete this thesis

successfully. My special thanks go to Dr. Ayman H. El-Hag for supporting me all the

times; he was always offering all kinds of technical and spiritual support. He taught

me dedication and loyalty to believe in every single part of my research. Above all, he

educated me with the invaluable mores and morals of conducting ethical and

professional research and enlightened my way to keep me in the right track. I want to

express my deep appreciation to Dr. Khaled Assaleh who was always advising me

with the trustworthy and truthful opinions in both my career and educational life. His

technical instructions were efficient and useful for continuing this research in the most

appropriate and professional way. My thankfulness goes to all my professors in the

electrical engineering department and colleagues at the American University of

Sharjah for their strong help and care throughout my Master’s degree study. I cannot

forget Eng. Fatima Al Dhaheri, Al Ain region maintenance department manager in

Abu Dhabi Transmission and Despatch Company for offering the appropriate research

conduction atmosphere and supplying me with the extremely useful data base I used

in my research. My special recognition is expressed to Eng. Fathi Abdulsalam my

senior engineer who was an elder brother, a helping senior and a trustworthy friend

throughout my study. He offered me the best he could to conduct and come up with

solid research study. His pieces of advice not only helped me in my study, but also

gave me the right meaning of professional life and how to deal with it. Last but not

least, my special gratefulness goes to my parents, those who worked in the dark and

was always available for me with the best they could to support me all the times until

I completed my work successfully. Their warm and frank instructions were always the

reason for my success in my life.

CHAPTER 1

INTRODUCTION

1.1 Transformer Asset Management

Transformer life cycle management has been the main concern to asset experts.

The main objective behind implementing transformer asset management (AM) plans

is reducing the operational costs of the transformer in service and extending its

utilization period economically. Condition assessment and maintenance are the main

elements of applying transformer AM. Condition assessment requires implementing

condition monitoring and diagnostic and end-of-life assessment techniques. Based on

the transformer condition, different maintenance strategies are assigned. Applying

cost-effective and sufficient transformer condition assessment and maintenance

processes leads to obtaining optimum transformer AM.

Several transformer AM strategies have been proposed. An intelligent system is

proposed using fuzzy logic modeling to support an optimum and reliable transformer

management decisions, like replacement, refurbishment and relocation [22]. The

model depends on transformer criticalities, rate of aging, and remnant life as its main

inputs. The model has the potential to provide accurate assessment; however,

implementing it acquires major costs. The model suggests using large number of

advanced CM techniques, which require high conduction and interpretation expenses

in terms of money and expertise. Another approach is a condition-based AM strategy

utilizing standard diagnosis of CM techniques. The standard diagnosis methods have

been utilized to derive a failure probability model in terms of the electrical,

mechanical, thermal stresses and life time [2]. This model has not specified a well

1

established set of CM techniques, which reduces the opportunities of practical

implementations for this model.

Transformer health index is determined using inspections tests on transformers

[3]. The health index method recommended is found to be useful for determining the

transformer probability of failure and calculating its remaining life. However,

performing health index modeling has difficulties, especially when lack of organized

data base of the transformer condition occurs.

It is concluded that implementing reliable and accurate CM techniques, practicing

optimized maintenance plans and estimating the transformer remnant life accurately

leads to applying developed asset management strategies [4].

1.2 Transformer maintenance plans

Currently, the most common maintenance procedures followed in utilities are

corrective maintenance, preventive maintenance and reliability centered maintenance.

These three strategies represent the three main generations of maintenance starting

from the first generation, which is corrective maintenance and passing by the

preventive maintenance strategy until reaching recently reliability-centered

maintenance (RCM).

1.2.1 Corrective Maintenance

Corrective maintenance is basically conducted whenever an alarm or trip occurs

on the equipment causing an unplanned outage. Recently, corrective-based

maintenance strategy on a transformer is considered as a poor approach. This is

because corrective action on a transformer can have a cost that is equal to its

replacement. For example, refurbishing a tripped transformer requires a lot of testing

and inspection time, especially if an internal fault takes place. Therefore, long lasting

transformer outage cost is the main concerning consequence of implementing

corrective maintenance on a power transformer. However, transformer corrective

maintenance can decrease costs for certain faults that are not crucially affecting the

transformer operating status. In general, corrective maintenance is a useful approach

only if it is accompanying other maintenance strategies.

2

1.2.2 Preventive Maintenance

Preventive maintenance is considered as the most safe maintenance strategy as

incipient faults can be prevented by implementing preventive maintenance. Applying

preventive maintenance procedures depends on conducting comprehensive condition

monitoring and diagnostics activities on the transformer’s parts, like internal

insulation, external insulation and control circuits. The time span between conducting

these tests is selected based on the type of preventive maintenance technique used by

the utility.

Two different types of preventive maintenance are practiced by the utility. The

first approach is Time-Based Maintenance (TBM). TBM is implemented when the

time span between the transformer testing and inspections is fixed. Time based

maintenance has been the most common applied maintenance strategy among utilities

before raising the AM thinking. Taking into consideration transformer reliability into

service, time based maintenance reduces the failure risks immensely, especially if the

time span between the inspections of the transformer is short. However, this approach

consumes huge manpower and planned outage costs. Moreover, transformer faults

and unplanned outages may occur during the time between the preventive

maintenance inspections; particularly when the time span between inspections is long.

The second approach is Condition-Based Maintenance (CBM). CBM is the most

acceptable form of preventive maintenance as it cuts down manpower costs. This type

of preventive maintenance mainly focuses on applying detailed and comprehensive

Condition Monitoring (CM) and diagnostic techniques to the transformer to form

clear and accurate assessment about its condition. Such accurate assessment helps the

asset manager to develop asset maintenance plans and assign optimized time spans

between maintenance activities. Online CM of a transformer is a vital element of

implementing CBM to detect incipient faults in early stages. So, in cases of faults

detections, CBM reduces the repair costs in terms of spare parts, expertise and time.

However, online CM has large conduction cost. Besides, interpreting condition

monitoring techniques and scheduling effective time spans between maintenance

activities are not easy tasks to be performed.

3

1.2.3 Reliability-Centred Maintenance

Reliability-centered maintenance (RCM) is found to be the most optimum

maintenance strategy that combines the advantages of different maintenance strategies

and merges them in a comprehensive and optimized form. RCM can be called risk-

based maintenance as maintenance activities are assigned depending on the degree of

risk of the transformer in service. The degree of risk depends on two factors: the

probability of failure and its severity. The probability of failure of the transformer

depends on the transformer operating conditions like the rate of aging, loading and

service age. On the other hand, the degree of risk is determined by analyzing the

effect of the transformer under study on the reliability of the network. Cost wise,

RCM is efficient as it keeps the transformer operating performance equals to the

targeted transformer performance level.

In general, transformer maintenance plan is a key stage of implementing

transformer life cycle management since it improves transformer operational costs,

performance and risks. Table 1 shows contribution of each maintenance strategy in

improving transformer operating parameters.

1.3 Transformer Condition Assessment

Transformer condition assessment has become important as a result of shifting

the maintenance strategy to be conditionally-based. To serve as a solid backbone for

the transformer asset management plans, condition assessment process has to be

comprehensive, accurate and cost-effective. The application of these conditions leads

to optimum and reliable diagnosis of the transformer condition in service and

facilitates the failure prediction process. Transformer condition assessment has two

main stages: condition monitoring and diagnostic of certain parameters and end-of-

life assessment. It has been accepted that insulation condition assessment of the power

transformer is the most efficient approach towards achieving efficient and reliable

transformer condition assessment. The reason is that most of the transformer failures

occur due to insulation mechanical, electrical and thermal deteriorations.

4

Table 1: Maintenance techniques effect on cost, performance and risk

Maintenance

Strategy Costs Performance Risks

Corrective

Maintenance

• Low manpower costs

• High outage costs if

the failure requires

comprehensive

inspection

• Sometimes, high spare

parts costs

• operational

performance level

<required

performance level

• Highest risk:

direct and

significant

damages

Preventive

Maintenance

• Highest manpower

costs

• Low spare parts cots

• operational

performance level

>required

performance level

• Least risks:

failures are

prevented using

routinely checks

Reliability-

Centered

Maintenance(RCM)

• Optimized manpower

costs

• Optimized spare parts

costs

• optimized

performance:

operational

performance level

=required

performance level

• Optimized risks:

maintenance

priority is

assigned based on

the degree of risk

 

 

 

 

5

Several CM and diagnostic techniques have been developed to inspect the

transformer condition. Power transformer useful operating life has been defined as the

remaining life of the transformer winding paper insulation [1]- [8]. Therefore, the

importance of solid insulation condition monitoring and diagnostic methods has

increased.

Updated and reliable end-of-life assessment is necessary for maintaining

optimum transformer life cycle management. Critical decision like transformer

replacement, relocation and refurbishment are based on transformer remnant life

estimation.

1.4 Transformer Insulation System

1.4.1 Transformer oil

Transformer oil is produced as a byproduct of the distillation of crude oil.

Transformer oil can be paraffinic oil or Naphthenic oil. The paraffinic oil is composed

of alkanes, whereas the naphthenic is composed of cycloalkanes. Naphthenic oil is

preferred to be used among utilities as it is superior in gas absorbing tendency and

low-temperature applications. Transformer oil is mainly installed in oil-filled

transformers as a cooling medium; however, it is considered as reliable insulating

material inside the transformer. The quality of oil as a cooling medium and an

insulating material in transformers is measured using different electrical, chemical

and physical tests. The tests that are related to the research work done is mentioned in

this thesis.

Transformer oil is degraded as a result of thermal and electrical internal

transformer faults. These faults release energy that is capable of decomposing

transformer oil into different gases , which are hydrogen(H2), methane(CH4),

Ethane(C2H4), Ethylene(CH6), acetylene(C2H2), carbon monoxide (CO) and carbon

dioxide (CO2). Transformer oil can decompose CO and CO2 as a result of its

oxidation.

6

1.4.2 Transformer Solid Insulation

Solid insulation material in power transformers is mostly made up of Kraft paper.

Kraft paper consists of cellulose, which is a natural linear polymer of cyclic β-D-

glucopyranosyl monomers, hemicellulose and lignin. The cellulosic chains are bonded

inter-molecularly and intra-molecularly by the hydrogen bonds via the hydroxyl

groups in a crystalline and amorphous regions forming at the end fibrils and fibers

that support the Kraft paper mechanically. These fibrils and fibers are linked by the

amorphous regions of hemicelluloses and lignin. The main failure cause of Kraft

paper insulation has been referred to the loss of its mechanical integrity due to the loss

of the hydrogen bonds and the deterioration of the fibers.

Oil-impregnated cellulose is degraded due to three main mechanisms, which are

hydrolytic degradation, oxidative degradation (pyrolysis mechanism) and thermal

degradation.

Hydrolytic degradation is activated by the presence of hydrogen ions separated

from acids. Acid-hydrolysis breaks the hydrogen bonds in the cellulosic chains

yielding the glucose rings. This mechanism is followed by dehydration reactions

yielding furan derivatives. Furfuraldehyde, which is a stable compound to acids, is

produced from Xylan, while cellulose yields hydroxylmethylfurfural. The later is not

stable to acids, thus laevulinic acid and formic acid are formed. This process of acid

catalyzed hydrolysis produces a sum of two water molecules, which cause carboxylic

acids to dissociate yielding hydrogen atoms [9].Therefore, Acids and water content

can be regarded as accelerators for cellulose paper degradation. Several studies have

reported the accelerative effect of water in reducing the cellulosic insulation life [9]-

[11].

Pyrolysis degradation consists of the oxidation of the hydroxyl groups to form the

carbonyl and the carboxyl groups permitting the chain scission of cellulose.

At temperature less than 200 ̊C cellulose insulation is thermally degraded

producing glucose molecules, moisture, carbon oxides and organic acids, while

different thermal mechanism applies for temperature more than 200 ̊C.

7

1.5 Transformer Insulation Assessment Techniques

There are several well understood and common insulation testing and diagnostic

techniques practiced as maintenance procedures in power utilities. Insulation

monitoring and diagnostic techniques that are related to the conducted research in this

thesis work are mentioned in this chapter.

1.5.1 Insulation resistance test (Megger test)

Insulation resistance (IR) test is one of the most classical routine diagnostic tests

performed on the power transformer. Usually, it is conducted in routine time-based

maintenance along with oil quality tests and dissolved-gas-in-oil-analysis (DGA) to

monitor the transformer insulation system and construct an informative and efficient

trend analysis. Moreover, IR test plays an important role in the cases of unplanned

transformer outages’ activities as it contributes to the verification of the occurrence of

an internal fault or severe deterioration in the insulation system. IR has certain

drawbacks affecting its diagnosis efficiency. The measurement of IR is affected by

the contamination and temperature of the insulation system. Temperature has a major

influence on the IR measurement, especially in oil filled power transformers. The

insulation system in oil filled transformers consists of insulating oil and solid paper

insulation where their insulation resistances are affected by temperature

independently. So, insulating oil temperature should be noted when conducting

megger test. Contamination of the insulating oil can reduce the value of the IR;

therefore, oil break down voltage (BDV) and water content tests are conducted along

with IR in corrective maintenance process to verify the presence of contamination in

the insulation system. Moreover, IR is not efficient in detecting incipient faults that

occur due to the operation of the Buchholz relay. The response of the IR as an

insulation quality parameter to the presence of incipient faults is slower than the

Buchholz relay. Accordingly, DGA is conducted along with IR test to detect incipient

faults in emergency unplanned outages. Oil BDV, water content and dissolved gases

are considered as important parameters conducted in corrective maintenance activities

for inspecting and diagnosing the tripped transformer.

8

Table 2: IEC 60422 recommended action limits for oil BDV and water content testing [12]

Property Category

Recommended

Action Limit

Poor

Breakdown Voltage

170-400kV

<50

72.5 -170kV

<40

<72.5kV <30

Water Content

170-400kV

>10

72.5 -170kV

>15

<72.5kV >25

Acidity

>72.5kV >0.15

<72.5kV >0.13

1.5.2 Oil quality tests

Breakdown voltage: Oil BDV measurement gives an indication about the

presence of contamination resulting from degradation reactions and moisture in the

oil. However, BDV has a slower response to contamination particles when compared

to other oil quality parameter. This imposes the need to conduct other maintenance

checks to the oil like water content and interfacial tension.

9

Water content: Water content is a reliable parameter as it can measure the

transformer insulation system deterioration state. Water content is considered as both

an accelerator for the aging process and as an aging product. Water moves from

transformer paper insulation to the dissolved form in oil as oil temperature increases

and travels back to paper insulation as temperature decreases to achieve water

equilibrium in the oil-paper system. Therefore, an accurate diagnosis of the presence

of moisture in transformer oil requires a temperature effect correction.

Acidity (Neutralization number): Acids in transformer oil are considered as

accelerators of paper insulation aging through hydrolytic degradation. Table 2 shows

the recommended action limit for BDV, water content and acidity as per IEC 60422

standard. The standard basically specifies for maintenance engineers the oil quality of

transformers with different ratings by referring to the measurement of oil BDV, water

content and acidity.

1.5.3 Dissolved gas-in-oil analysis (DGA)

Dissolved gas-in-oil analysis is one of the most acceptable methods used to detect

internal incipient faults in the power transformer. Internal faults cause the

hydrocarbon oil and cellulose paper in power transformer to emit different gases like,

H2, CH4, C2H4, CH6, C2H2, CO and CO2. Depending on the concentrations of these

gases, the type of fault and its severity is detected. DGA has great diagnosis efficiency

as it is used to anticipate transformer incipient faults. According to IEEE C57.104

standard, actions can be taken based on the transformer oil Total Dissolved

Combustible Gases (TDCG) as shown in Table 3. The standard recommends the oil

sampling intervals and operating procedures along with stating the transformer

condition based on the measurement of the TDCG.

CO2/CO ratio has been accepted as prominent parameter that indicates paper

insulation involvement in transformer incipient faults. According to IEC 60599

standards, a value of less than three of the CO2/CO ratio indicates paper insulation

involvement in an internal transformer fault. Merging the two parameters, TDCG and

CO2/CO ratio, CO2 can be approximated to get the total combustible gases (TCG),

which serves as a maintenance planning and diagnosis tool. Table 4 shows the

Westinghouse standard for planning maintenance checks based on the concentration

of TCG.

10

However, DGA does not show efficiency in predicting aging and providing

general view about the cellulose insulation degradation. Except, CO and CO2, all

dissolved hydrocarbon gases in transformer oil are emitted due to the deterioration of

oil only.

Table 3: Actions based on TDCG according to c57.104 [13]

TDCG Levels (ppm)

TDCG rates

(ppm/day) Sampling Interval Operation Procedure

Condition 4 >4630

>30 Daily Consider removal of service

10—30 Daily Advise Manufacturer

<10 Weekly Exercise extreme caution.

Analyze for individual gases. Plan outage. Advise manufacturer

Exercise extreme caution. Plan outage

Condition 3 1921-4630

>30 Weekly

10--30 Weekly Analyze for individual gases

<10 Monthly Advise Manufacturer

Condition 2 721-1920

>30 Monthly Exercise extreme caution. Plan outage

10--30 Monthly Analyze for individual gases

<10 Quarterly Advise manufacturer

Condition 1 ≤720

>30 Monthly Exercise extreme Caution. Analyze for individual gases. Determine load

dependence

10--30 Quarterly Exercise extreme Caution.

<10 Annually Analyze for individual gases. Determine load dependence.

Continue a normal operation

11

Table 4: Westinghouse guidelines on TCG [13]

TOTAL COMBUSTIBLE GASSES RECOMMENDED ACTION  

Normal Aging Analyze again in 6-12 months

0-500 ppm  

 Decomposition maybe in excess of

normal aging Analyze again in 3 months

501-1200 ppm  

 

More than normal decomposition Analyze in one month

1201-2500 ppm  

 

2500 ppm and above Make weekly analysis

to determine gas generation rate contact manufacturer

 

 

1.5.4 Tensile Strength (TS)

Tensile strength has shown a sufficient capability to assess the cellulose material

degradation. Different life prediction models of power transformers have been

proposed using TS confirming that the end-of-life criterion for transformer reliable

operation is when its oil-impregnated cellulose paper reaches to 50% of its initial TS

value [16]. Another model is proposed in [15] showing that transformer paper

insulation can maintain its mechanical integrity until it reaches 20% of its initial TS

value using the wide-span and the zero-span measurements (tensile index).

1.5.5 Degree of Polymerization (DP)

Degree of polymerization, which is defined as the average length of the cellulosic

chain, measured by the average viscometric method is more convenient method than

TS and molecular weight for analyzing the Kraft paper aging [16]. It is found that DP

12

is strongly correlated with TS. The end-of-life criterion for power transformers in

service is determined to be around DP of 250 [10]. It is found that TS is constant

when DP is in the range of 500-950, then TS decrease directly with DP in the range of

200-500. DP is found to have a tan-sigmoid relationship with temperature. The

decrease in DP starts at 70 ̊C, which is the normal operating temperature for power

transformer, reaching the maximum rate of change at 100 ̊C until getting to the end-

of-life criteria for transformer solid insulation (DP of 200) at 150 ̊C [10]. DP has

been strongly correlated with paper contaminants and found to be the most useful

parameter for assessing the transformer solid insulation [17]. The solid insulation life

of the power transformer is predicted, with limited accuracy [16], by modeling the

kinetics of cellulose degradation using DP. It is proposed that ( ) is directly related

to aging time [9], [18]. However, a kinetic model derived by finding the average

molecular weight has showed that the rate of bond scission ( ) is not an accurate

method [16].

1.5.6 Oil furan content

All the mentioned parameters are credible in terms of assessing the paper

insulation degradation; however, they are destructive and costly to conduct. The

measurements taken to find the TS, DP and the molecular weight of cellulose

insulation paper should be conducted on paper samples taken from the transformers in

operation. This way of sampling requires transformers’ outages which is costly to the

utilities. Therefore, furans’ concentration in transformer oil can be a promising

indirect measurement to the aging of transformer paper insulation.

Furan content measurement consists of measuring six furfurals dissolved in

transformer oil that differs in their stability. Furfurals are composed of 2-furaldehyde

(furfural), 5-hydroxymethyl-2-furaldehyde, 2-furoic acid, furfuryl alcohol, 5-methyl-

2-furaldehyde and 2-acetyl furan compounds where 2-furaldehyde is the most stable

compound for measurement. Furans as degradation measurements of transformer

solid insulation have gained wide acceptance in utilities as it is more practical to

sample the oil online without the need to have a transformer outage. Furthermore,

furan content can be a complementary monitoring technique to DGA. Furans in

transformer oil have superiority to detect paper insulation degradation over oil quality

13

measurements and DGA as they are products of the degradation of only paper

insulation and have been correlated with other paper aging measurements.

Several researchers have introduced furans as efficient and practical tools for

transformer condition assessment and its life cycle estimation. A large scale survey

conducted on large population of mineral oil transformers is presented in [18]. It has

been found that certain thresholds for the furans’ concentrations in transformer oil

does exist and can rank the transformer operating state from healthy to poor. The

remaining transformer life was calculated using statistical method based on the

distribution of the transformer population. The study confirmed that the total furan

concentration in transformer oil is more reliable indicator of paper degradation than

each individual furan, although each individual furan compound can be useful for

indicating types of faults and their severities [19]. A comprehensive diagnostic

method is derived based on statistical correlations between transformer internal gases

and oil quality parameters and furfural [21]. Trend analysis is used in [21] utilizing

the correlation between DGA and furfural to diagnose transformer aging.

In [7], experience is utilized to interpret the furanic compounds concentrations

using two developed approaches, which are the percentile score approach and the

prediction of degree of polymerization (DP) approach. Further investigation has led to

forming a relationship between the concentration of furfural and DP [22]

1.5.7 Artificial Intelligence Applications in Transformer Condition Monitoring

and Diagnostic

Attempts have been conducted to predict transformer insulation quality

parameters. A successful prediction of BDV and interfacial tension has been

performed using a cascade configuration of artificial neural networks with IR

resistance as input feature vector for the used model [23]. Polynomial networks have

shown their prediction capability in the cases of shortage of number of training

samples to correlate BDV and interfacial tension with IR [24]. However, in [23]and

[24], the proposed models have failed to predict both transformer oil water content

and DGA. Therefore, these proposed models failed to offer comprehensive and cost-

effective CM monitoring and diagnostic technique. To be comprehensive and cost-

effective, especially in emergency transformer outages, prediction of BDV, water

content and dissolved gases should achieved. Therefore, by limiting the model to

predict BDV and interfacial tension, costs of conducting the other tests remain and the

14

required time for testing in outages situations is huge and costly. A reliable neural

network (NN) and polynomial regression models were presented successfully to

correlate BDV with oil acidity, oil water content and transformer age [25]- [26].

Except for few cases, the percentage error between the actual and predicted values of

transformer breakdown voltage was less than 10%. However, the model needs the

water content and total acidity as inputs to predict the breakdown voltage, which

makes the model relatively costly. Megger test is the preferred input for predicting

transformer oil quality parameters as it is one of the cheapest and easiest CM tests.

Besides, still prediction of transformer dissolved gases for detecting incipient faults is

not approached; therefore, proposed models in [25]- [26] are not cost-effective.

Predicting dissolved gases in transformer oil has been attempted recently to

facilitate the condition assessment process and reduce its costs. Support vector

machine method with genetic algorithm is used to forecast the transformer oil

dissolved gases [27]. The model has predicted different dissolved gasses with

maximum percentage error of 6.7%. Similarly to what has been presented in [26], the

model requires the history data for the dissolved gasses for each unit to be predicted.

Such requirement limits the generalization of the model. Trend of dissolved gases is

approximated using grey model, which functions efficiently when a shortage of

sampling data occurs [28]. The model proposed in [28] consists of two stages: the

forecasting of gases trends stage and incipient faults diagnosis stage using common

standards. The proposed grey model has achieved a total absolute mean percentage

error of 2.5%. However, from an asset management point of view, utilities prefer to

measure transformer oil dissolved gases and get more reliable and clear image about

the internal condition of the transformer rather than forecasting DGA. Most

importantly, the models proposed in [27] - [28] needs DGA as an input which requires

relatively big portion of time for conduction and expertise for analysis. This time is

considered extremely costly in emergency outages. DGA prediction approach can be

more of great value if it is utilized to reduce the unplanned outage time for a tripped

transformer. Self organizing polynomial networks, neural networks, expert systems

and fuzzy logic have been utilized to assess the condition of the transformer using

DGA [29]- [32]. Self-organizing polynomial networks have shown excellent

performance of classifying eight different types of transformer faults. The system has

achieved a classification accuracy of 98% for a dependable number of data base of

711 samples [29]. It is found that polynomial networks can enhance the diagnostic 15

efficiency of DGA. MLP ANN has proven its efficiency in diagnosing transformer

incipient faults using 150 data base of transformer oil dissolved gases [30]. In [30]

different input and output patterns are chosen from different diagnostic standards like

IEC and Rogers methods. The proposed ANN has achieved a diagnosing efficiency of

87-100% depending on the input and output patterns chosen. The tendency of

transformers’ incipient faults and the severity of them are quantitatively measured

using fuzzy logic technique [31]. In [32], a combination of multi-layer perceptron

(MLP) ANN with expert system has been utilized to classify transformer incipient

faults depending on the concentrations of the dissolved gases in transformer oil. The

expert system used consists of different modules that represent different types of

standards diagnosing different types of faults. The proposed system has sown a better

diagnostic accuracy (95%) than MLP ANN (93%) and expert system (89%) alone.

The condition assessment process is based on classifying the pattern of dissolved

gases in transformer oil to detect developed incipient faults. These models enhance

the diagnosis efficiency of DGA; however, diagnosis of transformer solid insulation is

not developed. In other words, these models can be more practical if they, in addition

to fault diagnosis, have a prediction stage for transformer aging. This solid insulation

assessment can improve condition assessment as end-of-life assessment is attained. In

addition to that, the previous reviewed models require the conduction of DGA. Cost

of DGA conduction is acceptable in preventive maintenance; however, much larger

outage cost is acquired for DGA in corrective maintenance.

Artificial neural networks have been proposed, using furfurals as prominent

inputs to approximate the aging degree of power transformers [33]. Relative aging

degree (RAD) of power transformer is calculated using inputs of paper and oil quality

parameters. The inputs to the ANN model used in this study are furfural, volume of

carbon monoxide (CO), Interfacial tension and acidity. Later study has proposed

ANN model with three inputs, which are CO concentration, carbon dioxide

concentration (CO2) and furfural [8] to predict the transformer age. The model is

verified using transformer samples in service with mean square error of 3.72%.

However, using furfural as an input is a costly approach to conduct by utilities.

16

1.5.8 Summary

It obvious from the review of the state-of-the-art research in transformer

condition assessment strategies using artificial intelligence that the two transformer

operational parameters, which are costs and performance, are not minimized in the

cases of the corrective maintenance and preventive maintenance. In corrective

maintenance procedures, some CM parameters are predicted successfully like BDV

and interfacial tension, but others like dissolved gases are not. Therefore, the

corrective maintenance conduction time and costs are high. In preventive

maintenance, accurate aging degree and solid insulation assessment have been

achieved; however, with relatively costly procedures. Accordingly, this research work

is concentrated on developing reliable transformer condition assessment strategies

through artificial intelligence, specifically ANN. The proposed condition assessment

strategies in this thesis should enhance the diagnosis efficiency of CM techniques and

minimize the transformer operational costs and performance in both preventive

maintenance and corrective maintenance activities.

17

CHAPTER 2

OBJECTIVES AND CONTRIBUTIONS OF THE RESEARCH

2.1 Objectives

The main objective in this research is to enhance the diagnosis efficiency of CM

and diagnostic techniques to provide reliable and cost-effective condition assessment

process. Well-organized condition assessment leads to implementing efficient

transformer maintenance plans and accordingly transformer AM with an optimized

operational costs, performance and risks. Two main studies are conducted in this

research, which focus on predicting transformer diagnostic parameters.

The first study concentrates on enhancing the diagnostic efficiency of transformer

insulation resistance test in corrective maintenance activities by predicting its

accompanying tests like BDV, water content, and dissolved gases using ANN. When

conducting a corrective maintenance activity, especially after the operation of the

Buckhoulz relay, differential relay and the restricted earth fault relay, successful

prediction of these tests can reduce their conduction and diagnostic lasting hours.

Therefore, the transformer internal inspection time and the unplanned outage costs of

conducting a corrective maintenance activity are reduced.

The second study focuses on achieving a cost-effective and reliable method

towards assessing transformer solid insulation. This target is attained by predicting

furan content in transformer oil using ANN. Classical, easy to conduct and well

understood tests’ measurements are used as inputs for an ANN-based prediction

model to approach a cost-effective transformer solid insulation assessment model.

Reliability is maintained by choosing furan content as a measuring parameter for

18

determining the transformer’s solid insulation useful life into service. Therefore, an

optimized end-of-life assessment strategy can be accomplished.

2.2 Contribution and Significance of the Research

Classical transformer testing methods, such as Insulation resistance and

transformer oil testing, have been recognized among utilities as the most practiced

methods. These methods are cost-effective and provide acceptable accuracy for

efficient condition assessment. Therefore, several standards are developed to

formalize the diagnosis of the insulation resistance and oil testing. Predicting

insulation testing parameters using classical testing methods can enhance the

diagnosis efficiency of the classical methods with optimizing the cost of maintenance

tests in preventive maintenance activities. Considering emergency outages of power

transformers, condition assessment of the tripped transformer is an elementary task

before reconnecting it to the network. Condition assessment of tripped transformer

requires certain tests like IR, BDV, oil water content and DGA. Each test has its own

conducting time depending on the availability of resources and the complexity of the

test. IR testing is relatively cost-effective in terms of conduction time and activity

cost. Therefore, predicting the other tests conducted in the case of unplanned outage

using insulation resistance test can reduce the outage time and accordingly the outage

cost. In comparing this prediction model with the other prediction models proposed in

the most recent research, the proposed model is unique in terms of providing

comprehensive assessment through prediction of the important needed parameters

rather than predicting some of them for taking a reconnection decision in corrective

maintenance procedure. Accordingly, costs are reduced in terms of unplanned outage

time, manpower costs and testing costs.

Furan content in transformer oil can enhance DGA diagnosis by indicating

cellulose paper deterioration. However, it is not cost-effective to conduct furan

content for each oil sample taken from the transformer. Therefore, predicting furan

content in transformer oil can enhance the implementation of cost-effective preventive

maintenance strategy. Moreover, keeping a track of predicted furan content in

transformer oil can give a valuable indication about the remnant reliable life of the

solid insulation of the power transformer. As it improves transformers’ operational

costs and performance, the proposed furan content prediction model is a novel model 19

that can be considered superior to the state of the art models for assessing transformer

solid insulation. Predicting furan content as per the proposed methodology utilizes

relatively cheap testing inputs to the model which improves the operational

performance and costs of preventive maintenance.

2.3 Related Publications

1. Refat Atef Ghunem, Ayman H. El Hag, Khaled Assaleh, and Fatima Al Dhaheri, “Prediction of Furan Content in Transformer Oil Using ANN,” Paper is accepted to be orally presented in International Symposium in Electrical Insulation, ISEI, June 6-9, 2010.

2. Refat Atef Ghunem, Ayman H. El Hag, and Khaled Assaleh, “Estimating Transformer Oil Parameters Using Artificial Neural Networks,” International Conference on Electric Power and Energy Conversion Systems, EPECS, November 10-12,2009.

More publications are under preparation to be submitted to IEEE Transactions

and IEEE Magazines.

 

 

 

 

 

 

 

 

 

 

 

 

 

20

 

   

CHAPTER 3

MATERIALS AND METHODS

3.1 Transformer Insulation Tests

The transformers under study belong to Abu Dhabi Transmission and Despatch

Company. The transformers’ service ages vary between 7-15 years. For the first

problem under study, which is enhancing the diagnosis efficiency of insulation

resistance parameter using ANN, nineteen power transformers are used. The

transformers have ratings between 50-140MVA and voltage rating of 220/33/11KV.

For the second research study, which is prediction of furan content of transformer oil

using ANN, forty transformers are used. The training and testing data used in the

model are collected from the transformers’ maintenance records for transformers with

voltage ratings of 220/33 KV, 132/11 KV, 380/15 KV and 33/11 KV. The oil used in

the tested transformers is mineral naphthenic oil.

3.1.1 Megger Test

The IR test is conducted by injecting a 2KV DC voltage into the transformer

insulation and the insulation resistances between high voltage (HV) terminal and

grounded main tank, low voltage (LV) terminal and grounded tank and tertiary

voltage (TV) terminal and grounded tank are determined. The measurement is taken

after 15 and 60 seconds until the reading is stabilized; otherwise, the reading is taken

after 300 seconds. Figure 1 shows the equivalent schematic diagram of the megger

test for HV/ ground measurement. The practical megger test conducted in the site is

shown in figure 2.

21

 

Figure 1: Schematic diagram of Megger test (HV/Ground)

 

Figure 2: Side view of practical megger test conducted in the site

22

3.1.2 Oil Quality Tests

The breakdown voltage of the transformer oil sample is measured by applying a

gradually increasing voltage between two uniform electrodes with a 2.5mm separation

gap as per IEC60156. IEC 60814 standard is used to measure the water content in the

transformer oil samples used in the study. The oil samples are taken at a temperature

range of 60-65 ºC to reduce the effect of water equilibrium in the oil/paper insulation

system. Acidity is measured in chemical lab according to ASTM D974 standard.

3.1.3 Dissolved-Gas-in-Oil Analysis (DGA)

Dissolved gas analysis (DGA) is conducted in the chemical lab using the

chromatographic method according to the IEC 60599 standard to find the

concentration of the following gases: Hydrogen (H2), Methane (CH4), Ethane (C2H4),

Ethylene (C2H6), Acetylene (C2H2), Carbon Oxide (CO), and carbon dioxide (CO2).

The total dissolved combustible gases (TDCG) are calculated by adding the

concentration of the measured gases except the concentration of CO2. Samples of IR,

oil testing and DGA results used in the first study are depicted in Table 5.

Table 5: Sample of the data used in the first study of predicting BDV, water content, TDCG and

CO2/CO ratio using insulation resistance measurements as input features

Insulation Resistance (MΩ) Parameters

HV/E MV/E TV/E

BDV

(kV)

Water Content

(ppm)

TDCG

(ppm) CO2/CO

300 240 400 71.4 26.8 497 8.25

820 880 1100 75 30.57 737 7.98

280 280 380 74 16.46 672 6.46

560 580 620 67 22.62 665 6.23

270 220 390 75 17.63 894 4.85

210 160 260 74.2 13.75 1126 4.17

 

23

3.1.4 Oil Furan Content

Furan content in transformer oil is measured using high performance liquid

chromatography technique according to ASTM D 1533 standard. Table 3 shows

samples of the oil testing data used in the second study. It is evident from the table

that the transformers oil results cover a wide range of transformer oil conditions.

Table 6: Samples of the oil data used the second study of predicting furan content in transformer oil

TR # CO

(ppm)

CO2

(ppm)

TCG

(ppm)

TDCG (ppm)

Water content (ppm)

Acidity (KOH/g)

BDV

(KV)

Furan content (ppm)

T1 393 4559 764 5323 39 0.54 28 17.96 T3 193 9128 322 9350 35 0.14 23 14.11 T9 298 2710 342 3052 21 0.19 54 4.23 T15 347 1816 361 2177 7 0.1 57 3.06 T16 229 1230 266 1496 13 0.11 60 2.81 T26 374 2836 577 3413 12 0.03 89 1.57 T36 58 1775 129 1904 14 0.04 58 0.04 T37 122 2386 190 3026 10 0.04 78 0.38 T40 66 1147 78 1225 9 0.02 77 0.19

3.2 Artificial Neural Networks (ANN)

Artificial neural networks (ANN) have been developed to be analogous with the

neural system in the human body with simple units called neurons. The neurons are

collecting signals from other neurons after being weighted in connection links. ANN

has proven their efficiency in different power system applications [34].

3.2.1 ANN Structure Proposed

A multi-layer perceptron (MLP) neural network is used to model the relationships

proposed in both studies. The use of ANN as a modeling technique has been proposed

due to the non linear and complex relationship between the inputs and the outputs

proposed. The problem of predicting insulation diagnostic parameters can be defined

physically, which makes it no linear and complex. No linear mapping has been found

for such prediction problems before. MLP ANN has proved its efficiency in mapping

between nonlinear and complex input and outputs variables. In addition to that, MLP

24

ANN has been successfully used in transformer insulation diagnosis and condition

assessment problems with high accuracy and reliability. Therefore, MLP ANN

topology has been chosen as a modeling technique in this research work. The first

studied ANN focuses on the relation between IR and oil BDV, water content and

dissolved gases. In this study the physical relationship between the insulation

resistance as an input and BDV, water content, TDCG and CO2/CO as outputs is

defined physically with an indirect approach. The indirect approach has come from

the fact that each parameter indicates different degradation mechanism or is able to

diagnose different insulation types. The variation in the diagnosis capability of the

parameters makes it more complex. The second studied ANN concentrates on the

relation between oil quality parameters and dissolved gases as inputs and furan

content in transformer oil as output. Similar to the first correlation problem, the

relationship between the oil quality parameters and furan content is defined physically

and in indirect way. The fact that DGA is an incipient fault diagnosis technique, oil

quality parameters are mainly oil degradation indicators and furan content is a paper

insulation assessment technique makes the correlation more complex to be derived. It

is a fact that a MLP ANN with one hidden layer can be efficient for any defined

correlation problem; however, more than one hidden layer can be efficient in some

highly complex and nonlinear problems. As the physical relationships proposed are

not defined directly, MLP ANN with two hidden layers for the nonlinear mapping can

be more efficient and useful. The need for using two hidden layers can be inferred by

examining [23]. The researchers have used a cascade of more than one MLP ANN

with one hidden layer to compensate for the indirect physical relations ships between

insulation resistance as an input and BDV, acidity and interfacial tension as outputs.

In this research, the two defined prediction problems have been experimentally tested

using MLP ANN with one hidden layer and different combinations of numbers of the

neurons. The achieved model has given low prediction accuracies for most of the

testing samples. Therefore, MLP ANN with two hidden layers has chosen to be the

modeling technique. Figures 2 and 3 show the MLP ANN architecture used for the

two prediction problems.

25

Figure 3 : Artificial neural network structure proposed for predicting BDV, water content, TDCG and

CO2/CO

 

 

Figure 4 Artificial neural network structure used for prediction of furan content

where is the dth feature, represents the weights between the input

unit and the hidden unit and is the weights between the output unit and

the hidden unit. The input samples in the first layer are sent to the hidden layer

through weighted connection links. The hidden layer calculates its net activation as in

the following equation:

26

1

Where is the number of features, which is three features in this study (HV/E,

LV/E and TV/E) , and represents the weights between the input unit and the

hidden unit. The output of the hidden layer, which is a nonlinear function of its net

activat nio , is shown as the following:

2

Where is the output of the hidden layer. The output layer calculates its net

activation as in the following equation:

3

Where is the number of hidden neurons and is the weights between the

input unit and the hidden unit. The output layer yields an output as a

nonline r oa function f its net activation as shown in the following equation:

4

where is the output at the output unit, which is equal to one in the studied

model.

The neural network used in the first model has three neurons in the input layer,

two hidden layers with three neurons at the first hidden layer and 10 neurons at the

second hidden layer and one neuron at the output layer. The number of the neurons in

the hidden layers has been chosen by tuning it until descent prediction accuracy is

achieved. Each of the three inputs, which are HV/E, LV/E and TV/E measured

insulation resistances, are presented to one of the input neurons. BDV, water content,

TDCG and CO2/CO ratio appear at the output neuron separately depending on the

needed predicted parameter. While the second model of prediction furan content in

transformer oil has input layer with different number of neurons depending on the

number of current inputs used, two hidden layers and one output layer for predicting

furan content as an outcome of the ANN. By tuning the number of neurons in the

hidden layers until reaching the highest prediction accuracy, five neurons at the first

hidden layer and ten neurons at the second hidden layer are found to form the most

optimum furan content prediction model.

27

Tan sigmoid activation function due to its non linearity to map the non linear

correlation between the proposed inputs and outputs. Since gradient descent is used,

tan sigmoid function is continuous as defined in equation 5.

,2

1 · ,1 5

The output of neural network can be expressed as a function of the inputs, the

weights between input layer and the hidden layer and the weights between the hidden

layer and the output layer as per the following equation:

6

3.2.2 Feature extraction using stepwise regression

Stepwise regression is a common feature extraction method that selects the model

inputs depending on their statistical significance. Stepwise regression can be of great

value when availability of redundant input features of the prediction model occurs,

which can cause a reduction in prediction efficiency. Applying stepwise regression

analysis requires conducting certain steps that contain measuring the statistical

significance of the model approached using the partial F-statistic parameter. Partial F-

stati ic sst i determined as shown in the following equation [35].

⁄ , , … , , , … , 7

where is the regression coefficient due to the current added term , , , … ,

, , … , are the regression coefficients for the terms currently in the model

indicates the mean square error for the model and ⁄ , , … , ,

, … , indicates the regression sum of squares due to given that , , … ,

, , … , are already in the model.

For this research work, stepwise regression is an important step for predicting

furan content in transformer oil. Stepwise regression can be efficient in testing the

possible input features of the model and accordingly an optimum feature modeling

step is achieved. Stepwise regression is not utilized in the first research study, which

is enhancing the diagnosis efficiency of IR. In a corrective maintenance situation,

megger test is conducted first for insulation inspection as it is easier and other

28

accompanying tests are conducted later. Therefore, practically, megger test results are

the only and the most optimum input features for the proposed model.

Matlab is used as a feature modeling environment for predicting furan content in

transformer oil. Terms are added or removed to the input feature vector depending on

the p-value of entering and removing input terms. The p-value represents the null-

hypothesis test probability with a tolerance of α for adding a term and a tolerance of β

for removing a tolerance. The process starts with an initial model with the first term

added. Usually, the first selected input term is the one that has the least p-value

provided that it is less than the entrance tolerance. After that, terms are added or

removed after calculating the p-value for each step [36]. For a term that is not in the

model, the calculated p-value represents the null hypothesis that the term added does

not contribute to the statistical significance of the model (has zero coefficient).

Similarly, for a term that is already in the model, the calculated p-value represents the

null hypothesis that the term does not contribute to the model (has zero coefficient).

Figure 5 clarifies the stepwise regression process. It should be mentioned that the

entrance and removal tolerances (α & β) are chosen before initiating the stepwise

process to be 0.1. Besides, it should be noted that different models with different input

features can be achieved depending on the first term added to the model. This

variation in the feature vectors reduces the generalization capability of the model and

makes it more localized for the sample of data used. Therefore, all combinations of

featuring models are analyzed by changing the first selected term each time until

using all the features nominated initially.

The statistical parameters used to compare between the analyzed models are

Adjusted R2-statisitc and F-statistic. R2-statisitc measures the proportion of variation

of the predicted variable (furan content in transformer oil) with respect to the selected

features in a multiple regression model. R2-statisitc that is equal to 1 represents the

best possible fit in a multiple regression model, while zero R2statisitc represents worst

fit between the input features and the predicted variable. R2-statistic can be calculated

as the following:

1 8

where SSR is the sum of squares regression and is calculated as in the following

equation:

29

Figure 5: Stepwise regression flow chart.

Y 9

SST is the sum of squares total. SST is calculated as the following formula:

Y 10

SSE is the sum of squares error. SSE is calculated as shown below:

30

11

where is the actual predicted furan content, Y is the mean of observations of

furan content used in the multiple regression model , is the predicted furan content

using the fitted model and N is the number of observations , which are transformer

oil samples in this study . However, R-square as a statistical parameter depends on the

number of the prediction variables used in the fitted model. This makes the

comparison between different models with different number of input variables unfair.

Therefore, Adjusted R2 parameter is used to compensate for the number of input terms

when comparing different multiple regression models with different predictors.

Adjusted-R2 is calculated as shown in the following equation:

11 1

1 12

where is the adjusted R2-statistic, n is the number of observations (transformer

oil samples) and p is the number of input variables in the multiple regression model.

F-statistic is a measure of the significance of the selected terms to predict the

dependent variable, which is furan content in transformer oil in this study. F-statistic

is calculated as following formula: in the

1 13

where SSR is the SST is the sum of squares regression, SSE is the sum of squares

error, p is number of variables and n is the number of observations (oil samples).

3.2.3 ANN Training and Testing

Before the training starts, the data is normalized by maximum value

normalization criterion to improve the accuracy of the model. The main objective

behind using the back-propagation training method is to use training samples of inputs

and outputs in the network to adjust the weights’ values ( , ) to minimize the

difference between the predicted and the actual outputs. The optimum weights

( , n: ) are learned by minimizing the training error given in the following equatio

, ∑ 14

Where , is the mean square error and is the target output at the kth

output unit. Using gradient descent, the updated weights are calculated as the

following:

31

15

16

Using the chain rule, and are calculated yielding the following

expressions:

17

18

In light of the limitation in the number of samples and to increase the statistical

significance of the results, round robin strategy is used in training and testing the

neural network. On a leave-k-out basis and cycling over all the samples, k samples of

the measured megger and oil analysis data are spared for testing while the rest of the

samples are used for training the neural network. Obviously as k increases the number

of test combinations increases while the size of the training samples decreases. For the

first study of enhancing the diagnosis efficiency of IR in obtaining our predication

results we started with k =1 and incremented its value by 1 up to 40% of the number

of measured samples. On the other hand, k is maintained to be one for the second

study of predicting furan content in transformer oil as statistical significance is

attained using stepwise regression.

Table 7 shows the mean square errors and their respective number of epochs of

sample of the training of ANNs proposed towards predicting BDV, water content,

TDCG, CO2/CO and furan content respectively. The criteria chosen for stopping the

training of the ANNs are reaching zero mean square error or its global minimum until

it converges to a constant values.

32

Table 7: Mean square error and their respective number of epochs for samples of the ANNs proposed

to predict BDV, water content, TDCG, CO2/CO and furan content

Proposed prediction output Mean square error Number of epochs BDV 0.001 30

water content 0.01 13 TDCG 0.2 13

CO2/CO 0.006 30 Furan content 1×10-11 500

It is found that there is a decent mapping between the proposed inputs and

outputs is reached. The good mapping is verified using testing inputs for predicting

BDV, water content, TDCG, TCG and furan content and then calculating the

percentage prediction accuracy using the absolute percentage error (APE) and mean

absolute percentage errors (MAPE) criteria. The prediction accuracy of the ANN

studied are determined using APE and MAPE criteria as shown in the following

equations:

19 100

∑100 20

Where is the ith actual output and is the ith predicted output of the

network and N is the number of transformer samples.

33

CHAPTER 4

RESULTS AND DISCUSSION

4.1 Oil Quality Parameters and DGA Prediction Model

4.1.1 Oil BDV Prediction

The correlation between the measured values of the BDV with their predicted

values is shown in Figure 6. The average prediction accuracy of the BDV prediction

model using leave-1 out strategy is calculated to be 97%. No significant drop in the

accuracy of the predicted BDV was noticed when the number of tested samples

increased, Table 8. The average accuracy for BDV has only dropped 3% as depicted

in Table 8. The proposed BDV prediction model proves the strong correlation

between transformer IR and BDV. It is found that the obtained correlation in this

study is superior to the results obtained by the models proposed in [23] and [24]. The

correlation between IR and BDV comes from the fact that both transformer IR and oil

breakdown voltage give an indication about the dielectric capability of transformer oil

to withstand high electrical stresses. Although there is the narrow range of variations

of BDVs (63.6-78.8 kV) of transformers’ oil samples used in the training and testing

stages, the generalization capability of the proposed ANN applies. Having an oil

sample with a small BDV values (<50) can be considered as a rare case, especially in

power transformers. Therefore, the variation band in BDVs used in the study is

practically reliable.

34

Figure 6: Correlation between insulation resistance and oil BDV.

 

 

 

 

 

35

 

Table 8: Prediction accuracies of BDV for different sizes of training sets

Size of testing set (samples)

Prediction Accuracy

1 97

2 93

3 93

4 94

5 93

6 93

7 94

8 94

9 94

10 94

11 92

12 95

13 95

14 94

15 94

16 95

17 95

18 95

19 95

20 95

21 92

Average Prediction Accuracy

94

36

4.1.2 Oil water content prediction

For a training process with leave-1 out strategy, the average prediction accuracy

of water content is found to be 85% as shown in Figure 7. Water content in

transformer insulation system is recognized as one of the most deteriorating

contaminates in transformer oil. This gives an indication about the correlation

between oil water content and transformer insulation resistance as they both measure

the contaminated degree of the transformer insulation system.

Table 9 shows the prediction accuracies as a function of the size of the tested

sample and the average prediction accuracy has been found to be 80%. Similarly to

the BDV prediction, the reduction in the model efficiency due to the increase in the

testing sample set is not significant (less than 6%). However, compared to BDV the

prediction accuracy has been reduced between 10-15%. This can be attributed to the

relatively wider scatter of the data of water content (17-32 ppm) compared to BDV

(63.6-78.8 kV) and the relatively small size of the training set (31 vs. 54). Moreover,

the oil temperature effect in altering the value of the measured oil water content was

not completely considered as the oil sample was taken between 60-65 ºC. Such

narrow band of temperature variation may still influence the accuracy of the water

content measurement. The proposed model to predict oil water content in this study

can be considered superior to other proposed models [23] and [24]where the effect of

temperature variation on the measured samples was completely ignored.

Figure 7: Correlation between insulation resistance and oil water content.

37

Table 9: Prediction accuracies of oil water content for different sizes of training sets

Size of testing set (samples)

Prediction Accuracy

1 85

2 81

3 81

4 80

5 79

6 80

7 79

8 74

9 81

10 75

11 77

12 83

Average Prediction Accuracy 80

4.1.3 DGA prediction

a. Total Dissolved Combustible Gases (TDCG)

The correlation between TDCG and IR is verified in Figure 8 with a prediction

accuracy of 88%. Referring to Table 10, the average prediction accuracy for the

overall prediction process after increasing the testing sets’ sizes each time by one

sample is reduced to 74%. This average prediction rate shows that the correlation

between IR and TDCG does exist. Larger size of the training set can improve the

performance of the model immensely. TDCG has been accepted as a well informative

parameter to indicate the presence of incipient faults and degradation of the

transformer insulation system. The degradation process occurs due to the thermal,

38

electrical and mechanical stresses of the insulation materials. The degradation process

is accelerated by certain factors like, water content and acids. Internal faults can

accelerate the overheating of insulating materials causing the transformer oil and

cellulose paper insulation to degrade severely. Insulation resistance is affected by the

deterioration and the abnormal excess of degradation, thus, IR can be an indicative

parameter to the TDCG. Although TDCG is a reflection of incipient faults rather than

a parameter for aging index, an acceptable and reliable correlation between TDCG

and IR is verified in the proposed model.

Figure 8: Correlation between insulation resistance and oil TDCG.

 

 

 

 

 

 

 

 

 

 

39

 

Table 10: Prediction accuracies of oil TDCG for different sizes of training sets

Size of testing set (samples)

Prediction Accuracy

1 88

2 74

3 75

4 74

5 69

6 69

7 76

8 76

9 74

10 71

11 74

12 70

13 74

Average Prediction Accuracy

74

b. CO2/CO ratio

The correlation between CO2/CO ratio and transformer insulation resistance is

depicted in Figure 9. The prediction accuracy for the model in this case is calculated

to be 91%. However, Table 11 shows a relatively large reduction of the model

efficiency when the number of testing samples increases. The average prediction

accuracy after repeating the prediction process for all numbers of testing samples is

reduced to 78%. The reduction is attributed to the relatively small number of

measured samples used in training and testing the proposed model. CO2/CO ratio is a

40

good indicator of the cellulosic insulation involvement in internal fault or

deterioration of the transformer insulation. A CO2/CO ratio larger than seven can be

due to normal deterioration state of the transformer solid insulation. Both

deterioration process and incipient faults have their influence in decreasing the

insulation resistance. Although CO2/CO ratio is more efficient in diagnosing cellulose

involvement in incipient faults rather than indicating the deterioration of cellulose, the

proposed model shows an acceptable reliability of predicting CO2/CO ratio. This

agrees with the practice in utilities to use CO2/CO ratio with other insulation quality

parameters as indicators for cellulose thermal degradation.

   

 

Figure 9: Correlation between insulation resistance and oil CO2/CO ratio

41

Table 11: Prediction accuracies of oil CO2/CO ratio for different sizes of training sets

Size of testing set (samples)

Prediction Accuracy

1 91

2 82

3 76

4 74

5 77

6 84

7 76

8 67

9 80

10 82

11 80

12 75

13 74

Average Prediction Accuracy 78

4.2 Furan Content Prediction Model

Furan content in transformer oil is predicted using ANN by correlating furan

content with oil parameters contributing to paper insulation degradation. Different

cases are studied in this chapter depending on the input feature vector of the ANN.

Stepwise regression is used as a feature modeling method for improving the

prediction performance of the model.

42

4.2.1 Model Input Features

Seven different stepwise regression models with seven different initial terms are

statistically tested and analyzed. Matlab is utilized to compute the p-values of the

terms that are added or removed. Eventually, the optimum inputs are chosen as shown

in Table 12.

Table 12: Selected terms of different tested stepwise regression models

initial term selected terms

1. CO2 2. Acidity

3.BDV CO

1. CO2 2. Acidity

3.BDV CO2

1. CO2 2. Acidity

3.BDV TCG

1. CO 2. TDCG 3. Acidity 4. BDV

TDCG

1. CO2 2. Acidity

3.BDV water content

1. CO2 2. Acidity

3.BDV Acidity

1. CO2 2. Acidity

3.BDV BDV

43

According to the analyzed stepwise regression models, two different models are

recommended as shown in Table 12. The first model, which has CO, TDCG, BDV

and acidity as input terms, is recommended one time when the initial term is selected

to be TDCG. The second model, which has CO2, acidity and BDV as input features, is

suggested six times with CO2, acidity and BDV as initial features.

Analyzing the rate of occurrences of each of the two models recommended by

stepwise regression process, it is verified that the input features in second model are

the most statistically significant input terms for predicting furan content in

transformer oil. This is obvious when evaluating F-statistic to be the larger for the

second model as shown in Table 13.

Table 13: Recommended models by stepwise regression

Model inputs /Statistical Parameter

R-square Adjusted R-square F-statistic

CO, TDCG , BDV and Acidity

0.88 0.87 70.69

CO2, Acidity and BDV

0.88 0.87 98.77

Both of these stepwise regression models show that acidity is a crucial input

feature for predicting furan content in transformer oil. The correlation between furan

content and acidity comes from the fact that acidity is the main cause of hydrolytic

degradation of transformer paper insulation. Hydrolytic degradation yields at the end

furan derivatives in transformer oil. Stepwise regression model suggests that CO2 in

transformer oil can be an efficient indicator for transformer paper insulation

degradation. This is justified as CO2 is one of the main resultants of thermal and

hydrolytic degradation of transformer paper insulation. CO2 is found more efficient in

assessing the paper insulation in transformer oil than CO in this study. This result

44

verifies what has been found in [16] where it is concluded that CO has weak

correlation with paper aging. Moreover, in [19], [20] and [37] CO2 is found to be

more statistically significant for assessing paper insulation deterioration than CO. In

this study, CO as an input term is found to contribute to prediction of furan content in

transformer oil only when TDCG is used as an initial term in the model as shown in

Figure 4. CO2 is found to be correlated with furan content more than CO that is

mainly a resultant of thermal degradation. On the other hand, CO2 is a resultant of

both thermal and hydrolytic degradation, which is the main dominant aging

mechanism for transformer paper insulation under normal operating temperatures

[11].

BDV is found a statistically significant term for furan content prediction in

transformer oil. It is selected as in input feature in the model with highest statistical

significance along with CO2 and acidity. BDV is an efficient indicator of paper

insulation in transformer oil; BDV is reduced due to the presence of contamination

and particles resulted from the thermal, oxidative and hydrolytic reaction of

transformer paper insulation. The correlation between BDV and transformer paper

insulation aging is verified in [19].

When compared to CO2, BDV and acidity, the correlation between water content

and furan content in transformer oil is weak. At transformer operating temperatures,

water content in transformer oil is mainly produced due to the aging of transformer oil

as confirmed in [21]. Besides, water content in transformer oil is mainly produced as a

result of thermal degradation of paper insulation at temperatures higher than 100 ̊C

[16] and [38], whereas furan content measurements under study reflects aging of the

transformer at normal operating temperatures.

TCG are not found to be correlated with furan content in transformer oil. This

finding is confirmed in [19], where it is stated that TCG is not an efficient parameter

for assessing paper insulation degradation. Hydrocarbon gases like H2, CH4, C2H4,

C2H6 and C2H2 are composing both TDCG and TCG. Hydrocarbon gases are

indicators oil degradation; therefore, being part of TCG and TDCG may reduce the

statistical significance of TCG and TDCG to predict furan content in transformer oil.

So, the addition of CO to the second model is to improve the TDCG ability to

contribute to the prediction model. This is because hydrocarbon gases (H2, CH4, C2H4,

C2H6 and C2H2) that compose TDCG are adding redundancy to CO2, which is the

45

largest, portion composing TDCG. Therefore, it can be observed that CO can

contribute to predicting furan content in transformer oil, but with low significance. In

general it can be concluded from analyzing different stepwise regression models that

CO, CO2, BDV and acidity are efficient estimators of furan content in transformer oil.

4.2.2 Oil furan content prediction

The first model proposed in Table 13 is tested using ANN; the proposed input

model has a prediction accuracy of 74%. Figure 10 shows the predicted furan contents

for the 40 tested transformers along with their actual values. Generally, it can be said

that the model has an acceptable accuracy. It can be observed the prediction error is

higher in transformers with high concentrations of furan content (10-17 ppm) as

shown in Figure 10.

Figure 10: Furan content prediction in transformer oil with CO, TDCG, acidity and BDV as inputs

The second proposed model in Table 13 with CO2, acidity and BDV as input

features is tested using ANN as shown in Figure 11. The prediction accuracy is found

to be 85%. It is evident that this model has reduced the prediction accuracy, especially

for transformer with high concentrations of furan content (10-17 ppm). The results

found for the two models proposed in Table 13 using ANN verifies the results

obtained using stepwise regression.

46

Figure 11: Furan content prediction in transformer oil with CO2, acidity and BDV as inputs

CO is added to the input feature vector along with CO2, acidity and BDV to

analyze the effect of adding CO to the model. Figure 12 shows the actual furan

content in transformer oil for the 40 tested transformers along with their predicted

values; the prediction accuracy is found to be 90%. This models suggests that CO can

improve the performance of prediction model; however, with relatively small

significance.

Figure 12: Furan content prediction in transformer oil with CO, CO2, acidity and BDV as inputs

47

The fourth studied ANN prediction model contains CO, BDV and acidity as input

terms, while the fifth model has TCG, BDV and acidity as input features. The

redundancy effect of hydrocarbon gases are confirmed as the first model has

prediction accuracy of 73%, whereas the prediction accuracy of the second model is

69% as depicted in Figures 13 and 14. Moreover, by comparing the prediction

accuracies of the second model (input terms are CO2, acidity and BDV) and the fourth

model (input terms are CO, TDCG, acidity and BDV) it can be assured that CO2 is

more statistically significant that CO for predicting furan content in transformer oil

Figure 13: Furan content prediction in transformer oil with CO, acidity and BDV as inputs

 

 

 

 

 

 

 

 

48

 

 

 

 

 

 

 

 

Figure 14: Furan content prediction in transformer oil with TCG, acidity and BDV as inputs

49

CHAPTER 5

CONCLUSIONS AND RECOMMENDATIONS

5.1 Prediction of oil quality parameters and dissolved gases

In this thesis, an improvement to different condition monitoring and diagnostic

techniques conducted as maintenance procedures is achieved using ANN. ANN has

proved its efficiency to predict different transformer insulation diagnostic parameters

that are measured during maintenance. Oil BDV, water content, TDCG and CO2/CO

ratio are predicted with 97%, 85%, 88% and 91% accuracies respectively using ANN

as a prediction tool and transformer insulation resistance as input feature.

The proposed model of predicting BDV, water content, TDCG and CO2/CO ratio

to improve IR parameter shows the potential of AI techniques to improve transformer

insulation diagnosis and reduce the corrective maintenance costs. Prediction of

abnormal internal condition of transformers can be strengthened using the proposed

prediction strategy; however, larger size of data is required to have more reliable

model. The proposed model can be considered as a cost effective solution to the

disadvantage of implementing corrective maintenance on electrical equipment in the

system. Predicting transformer quality parameters reduces diagnosis time for the

transformer, thus, corrective action can be taken in an immediate manner.

Generalization of the model can be a critical issue; however, the proposed models in

this study suggests a decent potential for predicting oil BDV, water content, TDCG

and CO2/CO ratio. Relatively speaking, efficient and successful prediction is verified

for oil BDV and water content. Accepted prediction accuracy and decent

generalization capability for the TDCG prediction model has been proposed.

However, an improvement is suggested to oil CO2/CO ratio prediction model by

50

increasing the size of measured oil samples used in the training and testing process of

the ANN. In conclusion, the proposed correlation method can serve as an efficient

tool towards developing instructive transformer life management.

5.2 Prediction of Furan Content

Reliable remnant life estimation of transformer into service has been estimated

successfully using a cascade of two processes. The first process represents the feature

extraction method for selecting the significant inputs for the prediction model, which

is the stepwise regression. The second process composes of predicting the

concentration of transformer furan content using ANN. Accordingly, The correlation

between furan content and different transformer oil quality parameters has been

verified using the partial F-statistic parameter in a stepwise regression feature

extraction modeling process. Finally, a successful prediction of furan content in

transformer oil is achieved using ANN.

Stepwise regression is found to be a useful tool for extracting the optimum input

terms for predicting furan content in transformer oil. The reliability of stepwise

regression as an input modeling tool has been realized by verifying its outcomes with

conducted experiments mentioned in the literature. However, data with small size can

affect the efficiency of stepwise regression to be localized to the available

measurements. In addition to that, it should be noted that stepwise regression outcome

can differ with different selected initial terms. Therefore, to form a reliable conclusion

based on stepwise regression process, all possibilities of initial terms should be

included and the largest possible size of data should be used. In this study, ANN is

used after stepwise regression to enhancing the generalization of the proposed model

to predict furan content in transformer oil and to verify the results obtained from the

stepwise regression procedures.

The prediction accuracy of the first studied model with CO, TDCG, BDV and

acidity as inputs is 74%, while the prediction accuracy of the second studied model

with CO2, BDV and acidity as inputs is 85%. Therefore, the results obtained from

stepwise regression are verified using ANN. CO2 is proved to be more correlated

with transformer paper insulation aging than CO by comparing two ANN prediction

models. The first one has CO, acidity and BDV as inputs (prediction accuracy of

73%), while the second one has CO2, acidity and BDV as inputs (prediction accuracy 51

of 85%). The most efficient model is attained with a prediction accuracy of 90% when

CO, CO2, BDV and acidity as input features to the prediction model. It can be stated

that CO has small significance on predicting furan content as the prediction accuracy

of has increased by 5% only when adding CO to CO2, acidity and BDV inputs. Total

dissolved combustible gases are found to reduce the prediction accuracy to the

prediction model if added to the input feature vector. TCG contains hydrocarbon

gases that reflect incipient internal faults in transformer rather than transformer aging.

Predicting furan content using routine tests like BDV, water content, acidity and DGA

enhance the diagnostic efficiency of these tests. The proposed prediction model

improves these tests to be able to diagnose both oil and cellulose deterioration

efficiently and mange the transformer life cycle accordingly. Besides, the routine tests

used as input features of the model are easier to conduct and diagnose than furan

testing.

5.3 Recommendations for future work

The prediction models proposed can be tested on transformers belong to other

utilities in other countries with different operating conditions. This testing can verify

the generalization and validity of the proposed models. Predicting other insulation

quality parameters like interfacial tension and DP can improve the efficiency of the

proposed models for assessing the transformer condition. Applying different artificial

intelligence techniques like polynomial networks and comparing prediction accuracies

can be useful for achieving the most efficient technique.

52

REFERENCES

[1] M. Arshad, and S. M. Islam, “A Novel Fuzzy Logic Technique for Power

Transformer Asset Management,” 41st IAS Annual Industrial Meeting, pp. 276-

286, 2006.

[2] X. Zhang, and E. Gockenbach, “Asset Management of Transformers Based on

Condition Monitoring and Standard Diagnosis,” IEEE Electrical Insulation

Magazine, vol. 24, no. 4, pp. 26-40, 2008.

[3] N. Jahromi, R. Piercy, S. Cress, J. R.R. Service, and W. Fan, “An Approach to

Power Transformer Asset Management Using Health Index,” IEEE Electrical

Insulation Magazine, vol. 25, no. 2, pp. 20-34, 2009.

[4] A. E. B. Abu-Elanien, and M. M. A. Salama, “Asset management techniques for

transformers,” Electric Power System Research, vol. 80, no. 4, pp. 456-464,

April, 2010.

[5] J. Kachler, and I. Hohlein, "Aging of Cellulose at Transformer Service

Temperature. Part 1: Influence of Type of Oil and Air on the Degree of

Polymerization of Pressboard, Dissolved Gases, and Furanic Compounds in

Oil," IEEE Electrical Insulation Magazine, vol. 21, no. 2, pp. 15-21,

March/April, 2005.

[6] M. Ali, C. Eley, A. M. Emesley, R. Heywood, and X. Xaio, "Measuring and

Understanding the Aging of Kraft Insulating Paper in Power Transformers,"

IEEE Electrical Insulation Magazine, vol. 12, no. 3, pp. 28-34, May/June, 1996.

[7] R. D. Stebbins, D.S. Myers, and A. B. Shkolnik, “Furanic Compounds in

Dielectric Liquid Samples: Review and Update of Diagnosis Interpretation and

Estimation of Insulation Aging,” Proceedings of the 7th International

Conference on Properties and Applications of Dielectric Materials, pp. 921-

926, June, 2003.

[8] J. S. Navamany, and P. S. Ghosh, “Age Estimation of Cellulose Paper Insulation

in Power Transformers Using ANN,” Electrical Insulation Conference and

53

Electrical Manufacturing and coil Winding Technology Conference, pp. 277-

281, 2003.

[9] L. E. Lundgaard, W. Hansen, D. Linhjell, and T. J. Painter, “Aging of Oil-

Impregnated Paper in Power Transformers,” IEEE Transactions on Power

Delivery, vol. 19, no. 1, pp. 1-9, January, 2004.

[10] D. H. Shroff, and A.W. Stannett, “A Review of paper aging in power

transformers,” IEE Proceedings, vol. 132, no. 6, part C, pp. 312-319, November,

1985.

[11] I. Hohlen, and A. J. Kachler, “Aging of Cellulose at Transformer Service

Temperatures. Part 2. Influence of Moisture and Temperature on Degree of

Polymerization and Formation of Furanic Compounds in Free-Breathing

Systems,” IEEE Electrical Insulation Magazine, vol. 21, no. 5, pp. 20-24,

September/October, 2005.

[12] IEC standard 60422, Supervision and maintenance guide for mineral insulating

oil in electrical equipment, 3rd ed.

[13] IEEE std. C57.104-1991, IEEE Guide for the Interpretation of Gases Generated

in Oil-Immersed Transformers.

[14] "Protective Maintenance of Transformers by Gas-Oil Analysis" Westinghouse

Bulletin M7247.

[15] D. J. T. Hill, T. T. Lee, M. Darveniza, and T. Saha, “A study of degradation of

cellulosic insulation materials in power transformer. Part 2: tensile strength of

cellulosic insulation paper,” Polymer Degradation and Stability, vol. 49, no. 3,

pp. 429-435, 1995.

[16] M. Emsley, R. J. Heywood, M. Ali, and X. Xiao, “Degradation of cellulosic

insulation in power transformers. Part 4: Effects of aging on the tensile strength

of paper,” IEE Proceeding in Science, Measurement and Technology, vol. 147,

no. 6, pp. 285-290, November, 2000.

[17] D. J. T. Hill, T. T. Le, M. Darveniza, and T. Saha, “A Study of degradation of

cellulosic insulation materials in a power transformer. Part III: Molecular weight

study of cellulose insulation paper,” Polymer Degradation and Stability, vol. 51,

no. 2, pp. 211-218, February, 1996.

[18] M. Emsley, and G.C. Stevens, “Review of chemical indicators of degradation of

cellulosic electrical paper,” IEE Proceeding in Science, Measurement and

Technology, vol. 141, no. 5, September, 1994. 54

[19] J. R. Sans, K. M. Bilgin, and J. Kelly, “Large-scale survey of furanic

compounds in operating transformers and implications for estimating service

life,” Conference Record of the 1998 IEEE International Symposium on

Electrical Insulation, pp. 543-553, June, 1998.

[20] M. K. Pradhan, “Assessment of the Status of Insulation during Thermal Stress

Accelerated Experiments on Transformer Prototypes,” IEEE Transactions on

Dielectrics and Electrical Insulation, vol. 13, no. 1, February, 2006.

[21] M. Dong, G. J. Zhang, and Z. Yan, “Synthetic Furfural Analysis for

Transformer Aging,” Annual Report Conference on Electrical Insulation and

Dielectric Phenomena, CEIDP, pp. 245-247, 2004.

[22] B. Pahlavanpour, M. Eklund, M. A. Martins, “Insulating Paper Aging and

Furfural Formation,” Electrical Insulation Conference and Coil Winding

Technology Conference, pp. 283-288, 2003.

[23] K. Shaban, A. H. El-Hag, and A. Matveev, “A Cascade of Artificial Neural

Networks to Predict Transformers Oil Parameters,” IEEE Transactions on

Dielectrics and Electrical Insulation, vol. 16, no. 2, pp. 516-523, April, 2009.

[24] K. Assaleh, and A. H. El-Hag, “Estimating Transformer Oil Parameters Using

Polynomial Networks,” International Conference on Condition Monitoring and

Diagnosis, CMD, pp. 1335-1338, April, 2008.

[25] M. A. Wahab, “Artificial neural network-based prediction technique for

transformer oil breakdown voltage,” Electric Power System Research, vol. 71,

no. 1, pp. 73-84, September, 2004.

[26] M. A. Wahab, M. M. Hamada, A. G. Zeitoun, and G. Ismail, “Novel modeling

for the prediction of aged transformer oil characteristics,” Electric Power System

Research, vol. 51, no. 1, pp. 61-70, July, 1999.

[27] S. Fei, C. liu, and Y. Miao, “Support vector machine with genetic algorithm for

forecasting of key-gas ratios in oil-immersed transformer,” Expert Systems with

Applications, vol. 36, no. 3, part 2, pp. 6326-6331, April, 2009.

[28] S. Fei, and Y. Sun, “Forecasting dissolved gases content in power transformer

oil based on support vector machine with genetic algorithm,” Electric Power

System Research, vol. 78, no. 3, pp. 507-514, March, 2008.

[29] H. Yang, “Intelligent Decision Support for Diagnoses of Incipient Transformer

Faults Using Self-Organizing Polynomial Networks,” IEEE Transactions on

Power Systems, vol. 13, no. 3, August, 1998. 55

[30] J. L. Guardado, J. L. Naredo, P. Moreno, and C. R. Fuerto, “A comparative

Study of Neural Network Efficiency in Power Transformers Diagnosis Using

Dissolved Gas Analysis,” IEEE Transactions on Power Delivery, vol. 16, no. 4,

pp. 69-79, October, 2001.

[31] Q. Su, C. Mi, L. L. Lai, and P. Austin, “A Fuzzy Dissolved Gas Analysis

Method for the Diagnoses of Multiple Incipient Faults in a Transformer,” IEEE

Transactions on Power Systems, vol. 2 , no.2, pp. 344-348, May, 2000.

[32] Z. Wang, Y. liu, and P. J. Griffin, “A Combined ANN and Expert System Tool

for Transformer Fault Diagnosis,” IEEE Transactions on Power Delivery, vol.

13, no.4, pp. 1224-1229, October, 1998.

[33] I. N. Silva, A. N. Souza, R. M. C. Hossri, and J. H. C. Hossri, “Intelligent

System Applied in Diagnosis of Transformer Oil,” 8th international conference

on Dielectric Materials Measurements and Applications, pp. 330-334, 2000.

[34] R. Aggarwal, Y. Song, “Artificial Neural Networks in Power Systems III

Examples of Applications in Power Systems,” Power Engineering Journal, vol.

12, no. 6, pp. 279-287, 1998.

[35] T. Shanableh, and K. Assaleh, “Feature Modeling Using Polynomial Classifiers

and Stepwise Regression,” Neurocomputing, in press Accepted Manuscript,

March, 2010.

[36] Math works, Statistical Toolbox User Guide, 2008.

[37] M. Dong, Z. Yan, and G. J. Zhang, “Comprehensive diagnostic and aging

assessment method of solid insulation in transformers,” Annual Report

Conference on Electrical Insulation and Dielectric Phenomena, pp. 137-140,

2003.

[38] S. Levchik, J. Scheirs, G. Camino, W. Tumiatti, and M. Avidano,

“Depolymerization process in the thermal degradation of cellulosic paper

insulation in electrical transformers,” Polymer Degradation and Stability, vol.

61, 1998.

56

57

VITA

Refat Atef Ghunem was born in Syria in 1986. He was educated in government

schools in the United Arab Emirates (UAE) in Al-Ain. In 2004, he graduated from

high school holding the tenth rank among UAE government schools’ students with a

cumulative average of 99.2%. He received Abu Dhabi Water and Electricity

Authority (ADWEA) scholarship to join Electrical Engineering program at the

American University of Sharjah (AUS). Refat was awarded the Bachelor of Science

certificate in electrical engineering from AUS with Cum Laude standing in 2008.

In 2008, Mr. Refat began a master’s program in Electrical Engineering at AUS.

He was awarded the Master of Science degree in Electrical Engineering in 2010. Mr.

Refat has been working as electrical engineer in Abu Dhabi Transmission and

Despatch Company (TRANSCO) in the maintenance division since 2008.